1 /* 2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator 3 * 4 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to The DragonFly Project 7 * by Matthew Dillon <dillon@backplane.com> 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in 17 * the documentation and/or other materials provided with the 18 * distribution. 19 * 3. Neither the name of The DragonFly Project nor the names of its 20 * contributors may be used to endorse or promote products derived 21 * from this software without specific, prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.54 2008/07/01 02:02:54 dillon Exp $ 37 * 38 * This module implements a slab allocator drop-in replacement for the 39 * kernel malloc(). 40 * 41 * A slab allocator reserves a ZONE for each chunk size, then lays the 42 * chunks out in an array within the zone. Allocation and deallocation 43 * is nearly instantanious, and fragmentation/overhead losses are limited 44 * to a fixed worst-case amount. 45 * 46 * The downside of this slab implementation is in the chunk size 47 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu. 48 * In a kernel implementation all this memory will be physical so 49 * the zone size is adjusted downward on machines with less physical 50 * memory. The upside is that overhead is bounded... this is the *worst* 51 * case overhead. 52 * 53 * Slab management is done on a per-cpu basis and no locking or mutexes 54 * are required, only a critical section. When one cpu frees memory 55 * belonging to another cpu's slab manager an asynchronous IPI message 56 * will be queued to execute the operation. In addition, both the 57 * high level slab allocator and the low level zone allocator optimize 58 * M_ZERO requests, and the slab allocator does not have to pre initialize 59 * the linked list of chunks. 60 * 61 * XXX Balancing is needed between cpus. Balance will be handled through 62 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks. 63 * 64 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of 65 * the new zone should be restricted to M_USE_RESERVE requests only. 66 * 67 * Alloc Size Chunking Number of zones 68 * 0-127 8 16 69 * 128-255 16 8 70 * 256-511 32 8 71 * 512-1023 64 8 72 * 1024-2047 128 8 73 * 2048-4095 256 8 74 * 4096-8191 512 8 75 * 8192-16383 1024 8 76 * 16384-32767 2048 8 77 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383) 78 * 79 * Allocations >= ZoneLimit go directly to kmem. 80 * 81 * API REQUIREMENTS AND SIDE EFFECTS 82 * 83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we 84 * have remained compatible with the following API requirements: 85 * 86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty) 87 * + all power-of-2 sized allocations are power-of-2 aligned (twe) 88 * + malloc(0) is allowed and returns non-NULL (ahc driver) 89 * + ability to allocate arbitrarily large chunks of memory 90 */ 91 92 #include "opt_vm.h" 93 94 #include <sys/param.h> 95 #include <sys/systm.h> 96 #include <sys/kernel.h> 97 #include <sys/slaballoc.h> 98 #include <sys/mbuf.h> 99 #include <sys/vmmeter.h> 100 #include <sys/lock.h> 101 #include <sys/thread.h> 102 #include <sys/globaldata.h> 103 #include <sys/sysctl.h> 104 #include <sys/ktr.h> 105 106 #include <vm/vm.h> 107 #include <vm/vm_param.h> 108 #include <vm/vm_kern.h> 109 #include <vm/vm_extern.h> 110 #include <vm/vm_object.h> 111 #include <vm/pmap.h> 112 #include <vm/vm_map.h> 113 #include <vm/vm_page.h> 114 #include <vm/vm_pageout.h> 115 116 #include <machine/cpu.h> 117 118 #include <sys/thread2.h> 119 120 #define arysize(ary) (sizeof(ary)/sizeof((ary)[0])) 121 122 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x" 123 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \ 124 sizeof(int)) 125 126 #if !defined(KTR_MEMORY) 127 #define KTR_MEMORY KTR_ALL 128 #endif 129 KTR_INFO_MASTER(memory); 130 KTR_INFO(KTR_MEMORY, memory, malloc, 0, MEMORY_STRING, MEMORY_ARG_SIZE); 131 KTR_INFO(KTR_MEMORY, memory, free_zero, 1, MEMORY_STRING, MEMORY_ARG_SIZE); 132 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 2, MEMORY_STRING, MEMORY_ARG_SIZE); 133 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 3, MEMORY_STRING, MEMORY_ARG_SIZE); 134 KTR_INFO(KTR_MEMORY, memory, free_chunk, 4, MEMORY_STRING, MEMORY_ARG_SIZE); 135 #ifdef SMP 136 KTR_INFO(KTR_MEMORY, memory, free_request, 5, MEMORY_STRING, MEMORY_ARG_SIZE); 137 KTR_INFO(KTR_MEMORY, memory, free_remote, 6, MEMORY_STRING, MEMORY_ARG_SIZE); 138 #endif 139 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0); 140 KTR_INFO(KTR_MEMORY, memory, free_beg, 0, "free begin", 0); 141 KTR_INFO(KTR_MEMORY, memory, free_end, 0, "free end", 0); 142 143 #define logmemory(name, ptr, type, size, flags) \ 144 KTR_LOG(memory_ ## name, ptr, type, size, flags) 145 #define logmemory_quick(name) \ 146 KTR_LOG(memory_ ## name) 147 148 /* 149 * Fixed globals (not per-cpu) 150 */ 151 static int ZoneSize; 152 static int ZoneLimit; 153 static int ZonePageCount; 154 static int ZoneMask; 155 struct malloc_type *kmemstatistics; /* exported to vmstat */ 156 static struct kmemusage *kmemusage; 157 static int32_t weirdary[16]; 158 159 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags); 160 static void kmem_slab_free(void *ptr, vm_size_t bytes); 161 #if defined(INVARIANTS) 162 static void chunk_mark_allocated(SLZone *z, void *chunk); 163 static void chunk_mark_free(SLZone *z, void *chunk); 164 #endif 165 166 /* 167 * Misc constants. Note that allocations that are exact multiples of 168 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module. 169 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists. 170 */ 171 #define MIN_CHUNK_SIZE 8 /* in bytes */ 172 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1) 173 #define ZONE_RELS_THRESH 2 /* threshold number of zones */ 174 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK) 175 176 /* 177 * The WEIRD_ADDR is used as known text to copy into free objects to 178 * try to create deterministic failure cases if the data is accessed after 179 * free. 180 */ 181 #define WEIRD_ADDR 0xdeadc0de 182 #define MAX_COPY sizeof(weirdary) 183 #define ZERO_LENGTH_PTR ((void *)-8) 184 185 /* 186 * Misc global malloc buckets 187 */ 188 189 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches"); 190 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory"); 191 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers"); 192 193 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); 194 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery"); 195 196 /* 197 * Initialize the slab memory allocator. We have to choose a zone size based 198 * on available physical memory. We choose a zone side which is approximately 199 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of 200 * 128K. The zone size is limited to the bounds set in slaballoc.h 201 * (typically 32K min, 128K max). 202 */ 203 static void kmeminit(void *dummy); 204 205 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL) 206 207 #ifdef INVARIANTS 208 /* 209 * If enabled any memory allocated without M_ZERO is initialized to -1. 210 */ 211 static int use_malloc_pattern; 212 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW, 213 &use_malloc_pattern, 0, ""); 214 #endif 215 216 static void 217 kmeminit(void *dummy) 218 { 219 vm_poff_t limsize; 220 int usesize; 221 int i; 222 vm_pindex_t npg; 223 224 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 225 if (limsize > KvaSize) 226 limsize = KvaSize; 227 228 usesize = (int)(limsize / 1024); /* convert to KB */ 229 230 ZoneSize = ZALLOC_MIN_ZONE_SIZE; 231 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize) 232 ZoneSize <<= 1; 233 ZoneLimit = ZoneSize / 4; 234 if (ZoneLimit > ZALLOC_ZONE_LIMIT) 235 ZoneLimit = ZALLOC_ZONE_LIMIT; 236 ZoneMask = ZoneSize - 1; 237 ZonePageCount = ZoneSize / PAGE_SIZE; 238 239 npg = KvaSize / PAGE_SIZE; 240 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), 241 PAGE_SIZE, M_WAITOK|M_ZERO); 242 243 for (i = 0; i < arysize(weirdary); ++i) 244 weirdary[i] = WEIRD_ADDR; 245 246 if (bootverbose) 247 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024); 248 } 249 250 /* 251 * Initialize a malloc type tracking structure. 252 */ 253 void 254 malloc_init(void *data) 255 { 256 struct malloc_type *type = data; 257 vm_poff_t limsize; 258 259 if (type->ks_magic != M_MAGIC) 260 panic("malloc type lacks magic"); 261 262 if (type->ks_limit != 0) 263 return; 264 265 if (vmstats.v_page_count == 0) 266 panic("malloc_init not allowed before vm init"); 267 268 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 269 if (limsize > KvaSize) 270 limsize = KvaSize; 271 type->ks_limit = limsize / 10; 272 273 type->ks_next = kmemstatistics; 274 kmemstatistics = type; 275 } 276 277 void 278 malloc_uninit(void *data) 279 { 280 struct malloc_type *type = data; 281 struct malloc_type *t; 282 #ifdef INVARIANTS 283 int i; 284 long ttl; 285 #endif 286 287 if (type->ks_magic != M_MAGIC) 288 panic("malloc type lacks magic"); 289 290 if (vmstats.v_page_count == 0) 291 panic("malloc_uninit not allowed before vm init"); 292 293 if (type->ks_limit == 0) 294 panic("malloc_uninit on uninitialized type"); 295 296 #ifdef SMP 297 /* Make sure that all pending kfree()s are finished. */ 298 lwkt_synchronize_ipiqs("muninit"); 299 #endif 300 301 #ifdef INVARIANTS 302 /* 303 * memuse is only correct in aggregation. Due to memory being allocated 304 * on one cpu and freed on another individual array entries may be 305 * negative or positive (canceling each other out). 306 */ 307 for (i = ttl = 0; i < ncpus; ++i) 308 ttl += type->ks_memuse[i]; 309 if (ttl) { 310 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n", 311 ttl, type->ks_shortdesc, i); 312 } 313 #endif 314 if (type == kmemstatistics) { 315 kmemstatistics = type->ks_next; 316 } else { 317 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) { 318 if (t->ks_next == type) { 319 t->ks_next = type->ks_next; 320 break; 321 } 322 } 323 } 324 type->ks_next = NULL; 325 type->ks_limit = 0; 326 } 327 328 /* 329 * Calculate the zone index for the allocation request size and set the 330 * allocation request size to that particular zone's chunk size. 331 */ 332 static __inline int 333 zoneindex(unsigned long *bytes) 334 { 335 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */ 336 if (n < 128) { 337 *bytes = n = (n + 7) & ~7; 338 return(n / 8 - 1); /* 8 byte chunks, 16 zones */ 339 } 340 if (n < 256) { 341 *bytes = n = (n + 15) & ~15; 342 return(n / 16 + 7); 343 } 344 if (n < 8192) { 345 if (n < 512) { 346 *bytes = n = (n + 31) & ~31; 347 return(n / 32 + 15); 348 } 349 if (n < 1024) { 350 *bytes = n = (n + 63) & ~63; 351 return(n / 64 + 23); 352 } 353 if (n < 2048) { 354 *bytes = n = (n + 127) & ~127; 355 return(n / 128 + 31); 356 } 357 if (n < 4096) { 358 *bytes = n = (n + 255) & ~255; 359 return(n / 256 + 39); 360 } 361 *bytes = n = (n + 511) & ~511; 362 return(n / 512 + 47); 363 } 364 #if ZALLOC_ZONE_LIMIT > 8192 365 if (n < 16384) { 366 *bytes = n = (n + 1023) & ~1023; 367 return(n / 1024 + 55); 368 } 369 #endif 370 #if ZALLOC_ZONE_LIMIT > 16384 371 if (n < 32768) { 372 *bytes = n = (n + 2047) & ~2047; 373 return(n / 2048 + 63); 374 } 375 #endif 376 panic("Unexpected byte count %d", n); 377 return(0); 378 } 379 380 /* 381 * malloc() (SLAB ALLOCATOR) 382 * 383 * Allocate memory via the slab allocator. If the request is too large, 384 * or if it page-aligned beyond a certain size, we fall back to the 385 * KMEM subsystem. A SLAB tracking descriptor must be specified, use 386 * &SlabMisc if you don't care. 387 * 388 * M_RNOWAIT - don't block. 389 * M_NULLOK - return NULL instead of blocking. 390 * M_ZERO - zero the returned memory. 391 * M_USE_RESERVE - allow greater drawdown of the free list 392 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted 393 * 394 * MPSAFE 395 */ 396 397 void * 398 kmalloc(unsigned long size, struct malloc_type *type, int flags) 399 { 400 SLZone *z; 401 SLChunk *chunk; 402 SLGlobalData *slgd; 403 struct globaldata *gd; 404 int zi; 405 #ifdef INVARIANTS 406 int i; 407 #endif 408 409 logmemory_quick(malloc_beg); 410 gd = mycpu; 411 slgd = &gd->gd_slab; 412 413 /* 414 * XXX silly to have this in the critical path. 415 */ 416 if (type->ks_limit == 0) { 417 crit_enter(); 418 if (type->ks_limit == 0) 419 malloc_init(type); 420 crit_exit(); 421 } 422 ++type->ks_calls; 423 424 /* 425 * Handle the case where the limit is reached. Panic if we can't return 426 * NULL. The original malloc code looped, but this tended to 427 * simply deadlock the computer. 428 * 429 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used 430 * to determine if a more complete limit check should be done. The 431 * actual memory use is tracked via ks_memuse[cpu]. 432 */ 433 while (type->ks_loosememuse >= type->ks_limit) { 434 int i; 435 long ttl; 436 437 for (i = ttl = 0; i < ncpus; ++i) 438 ttl += type->ks_memuse[i]; 439 type->ks_loosememuse = ttl; /* not MP synchronized */ 440 if (ttl >= type->ks_limit) { 441 if (flags & M_NULLOK) { 442 logmemory(malloc, NULL, type, size, flags); 443 return(NULL); 444 } 445 panic("%s: malloc limit exceeded", type->ks_shortdesc); 446 } 447 } 448 449 /* 450 * Handle the degenerate size == 0 case. Yes, this does happen. 451 * Return a special pointer. This is to maintain compatibility with 452 * the original malloc implementation. Certain devices, such as the 453 * adaptec driver, not only allocate 0 bytes, they check for NULL and 454 * also realloc() later on. Joy. 455 */ 456 if (size == 0) { 457 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags); 458 return(ZERO_LENGTH_PTR); 459 } 460 461 /* 462 * Handle hysteresis from prior frees here in malloc(). We cannot 463 * safely manipulate the kernel_map in free() due to free() possibly 464 * being called via an IPI message or from sensitive interrupt code. 465 */ 466 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) { 467 crit_enter(); 468 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */ 469 z = slgd->FreeZones; 470 slgd->FreeZones = z->z_Next; 471 --slgd->NFreeZones; 472 kmem_slab_free(z, ZoneSize); /* may block */ 473 } 474 crit_exit(); 475 } 476 /* 477 * XXX handle oversized frees that were queued from free(). 478 */ 479 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) { 480 crit_enter(); 481 if ((z = slgd->FreeOvZones) != NULL) { 482 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); 483 slgd->FreeOvZones = z->z_Next; 484 kmem_slab_free(z, z->z_ChunkSize); /* may block */ 485 } 486 crit_exit(); 487 } 488 489 /* 490 * Handle large allocations directly. There should not be very many of 491 * these so performance is not a big issue. 492 * 493 * The backend allocator is pretty nasty on a SMP system. Use the 494 * slab allocator for one and two page-sized chunks even though we lose 495 * some efficiency. XXX maybe fix mmio and the elf loader instead. 496 */ 497 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) { 498 struct kmemusage *kup; 499 500 size = round_page(size); 501 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); 502 if (chunk == NULL) { 503 logmemory(malloc, NULL, type, size, flags); 504 return(NULL); 505 } 506 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ 507 flags |= M_PASSIVE_ZERO; 508 kup = btokup(chunk); 509 kup->ku_pagecnt = size / PAGE_SIZE; 510 kup->ku_cpu = gd->gd_cpuid; 511 crit_enter(); 512 goto done; 513 } 514 515 /* 516 * Attempt to allocate out of an existing zone. First try the free list, 517 * then allocate out of unallocated space. If we find a good zone move 518 * it to the head of the list so later allocations find it quickly 519 * (we might have thousands of zones in the list). 520 * 521 * Note: zoneindex() will panic of size is too large. 522 */ 523 zi = zoneindex(&size); 524 KKASSERT(zi < NZONES); 525 crit_enter(); 526 if ((z = slgd->ZoneAry[zi]) != NULL) { 527 KKASSERT(z->z_NFree > 0); 528 529 /* 530 * Remove us from the ZoneAry[] when we become empty 531 */ 532 if (--z->z_NFree == 0) { 533 slgd->ZoneAry[zi] = z->z_Next; 534 z->z_Next = NULL; 535 } 536 537 /* 538 * Locate a chunk in a free page. This attempts to localize 539 * reallocations into earlier pages without us having to sort 540 * the chunk list. A chunk may still overlap a page boundary. 541 */ 542 while (z->z_FirstFreePg < ZonePageCount) { 543 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) { 544 #ifdef DIAGNOSTIC 545 /* 546 * Diagnostic: c_Next is not total garbage. 547 */ 548 KKASSERT(chunk->c_Next == NULL || 549 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) == 550 ((intptr_t)chunk & IN_SAME_PAGE_MASK)); 551 #endif 552 #ifdef INVARIANTS 553 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd) 554 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount); 555 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart) 556 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount); 557 chunk_mark_allocated(z, chunk); 558 #endif 559 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next; 560 goto done; 561 } 562 ++z->z_FirstFreePg; 563 } 564 565 /* 566 * No chunks are available but NFree said we had some memory, so 567 * it must be available in the never-before-used-memory area 568 * governed by UIndex. The consequences are very serious if our zone 569 * got corrupted so we use an explicit panic rather then a KASSERT. 570 */ 571 if (z->z_UIndex + 1 != z->z_NMax) 572 z->z_UIndex = z->z_UIndex + 1; 573 else 574 z->z_UIndex = 0; 575 if (z->z_UIndex == z->z_UEndIndex) 576 panic("slaballoc: corrupted zone"); 577 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 578 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 579 flags &= ~M_ZERO; 580 flags |= M_PASSIVE_ZERO; 581 } 582 #if defined(INVARIANTS) 583 chunk_mark_allocated(z, chunk); 584 #endif 585 goto done; 586 } 587 588 /* 589 * If all zones are exhausted we need to allocate a new zone for this 590 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see 591 * UAlloc use above in regards to M_ZERO. Note that when we are reusing 592 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and 593 * we do not pre-zero it because we do not want to mess up the L1 cache. 594 * 595 * At least one subsystem, the tty code (see CROUND) expects power-of-2 596 * allocations to be power-of-2 aligned. We maintain compatibility by 597 * adjusting the base offset below. 598 */ 599 { 600 int off; 601 602 if ((z = slgd->FreeZones) != NULL) { 603 slgd->FreeZones = z->z_Next; 604 --slgd->NFreeZones; 605 bzero(z, sizeof(SLZone)); 606 z->z_Flags |= SLZF_UNOTZEROD; 607 } else { 608 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); 609 if (z == NULL) 610 goto fail; 611 } 612 613 /* 614 * How big is the base structure? 615 */ 616 #if defined(INVARIANTS) 617 /* 618 * Make room for z_Bitmap. An exact calculation is somewhat more 619 * complicated so don't make an exact calculation. 620 */ 621 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]); 622 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8); 623 #else 624 off = sizeof(SLZone); 625 #endif 626 627 /* 628 * Guarentee power-of-2 alignment for power-of-2-sized chunks. 629 * Otherwise just 8-byte align the data. 630 */ 631 if ((size | (size - 1)) + 1 == (size << 1)) 632 off = (off + size - 1) & ~(size - 1); 633 else 634 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK; 635 z->z_Magic = ZALLOC_SLAB_MAGIC; 636 z->z_ZoneIndex = zi; 637 z->z_NMax = (ZoneSize - off) / size; 638 z->z_NFree = z->z_NMax - 1; 639 z->z_BasePtr = (char *)z + off; 640 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; 641 z->z_ChunkSize = size; 642 z->z_FirstFreePg = ZonePageCount; 643 z->z_CpuGd = gd; 644 z->z_Cpu = gd->gd_cpuid; 645 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 646 z->z_Next = slgd->ZoneAry[zi]; 647 slgd->ZoneAry[zi] = z; 648 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 649 flags &= ~M_ZERO; /* already zero'd */ 650 flags |= M_PASSIVE_ZERO; 651 } 652 #if defined(INVARIANTS) 653 chunk_mark_allocated(z, chunk); 654 #endif 655 656 /* 657 * Slide the base index for initial allocations out of the next 658 * zone we create so we do not over-weight the lower part of the 659 * cpu memory caches. 660 */ 661 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) 662 & (ZALLOC_MAX_ZONE_SIZE - 1); 663 } 664 done: 665 ++type->ks_inuse[gd->gd_cpuid]; 666 type->ks_memuse[gd->gd_cpuid] += size; 667 type->ks_loosememuse += size; /* not MP synchronized */ 668 crit_exit(); 669 if (flags & M_ZERO) 670 bzero(chunk, size); 671 #ifdef INVARIANTS 672 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) { 673 if (use_malloc_pattern) { 674 for (i = 0; i < size; i += sizeof(int)) { 675 *(int *)((char *)chunk + i) = -1; 676 } 677 } 678 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ 679 } 680 #endif 681 logmemory(malloc, chunk, type, size, flags); 682 return(chunk); 683 fail: 684 crit_exit(); 685 logmemory(malloc, NULL, type, size, flags); 686 return(NULL); 687 } 688 689 /* 690 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE) 691 * 692 * Generally speaking this routine is not called very often and we do 693 * not attempt to optimize it beyond reusing the same pointer if the 694 * new size fits within the chunking of the old pointer's zone. 695 */ 696 void * 697 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) 698 { 699 SLZone *z; 700 void *nptr; 701 unsigned long osize; 702 703 KKASSERT((flags & M_ZERO) == 0); /* not supported */ 704 705 if (ptr == NULL || ptr == ZERO_LENGTH_PTR) 706 return(kmalloc(size, type, flags)); 707 if (size == 0) { 708 kfree(ptr, type); 709 return(NULL); 710 } 711 712 /* 713 * Handle oversized allocations. XXX we really should require that a 714 * size be passed to free() instead of this nonsense. 715 */ 716 { 717 struct kmemusage *kup; 718 719 kup = btokup(ptr); 720 if (kup->ku_pagecnt) { 721 osize = kup->ku_pagecnt << PAGE_SHIFT; 722 if (osize == round_page(size)) 723 return(ptr); 724 if ((nptr = kmalloc(size, type, flags)) == NULL) 725 return(NULL); 726 bcopy(ptr, nptr, min(size, osize)); 727 kfree(ptr, type); 728 return(nptr); 729 } 730 } 731 732 /* 733 * Get the original allocation's zone. If the new request winds up 734 * using the same chunk size we do not have to do anything. 735 */ 736 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 737 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 738 739 zoneindex(&size); 740 if (z->z_ChunkSize == size) 741 return(ptr); 742 743 /* 744 * Allocate memory for the new request size. Note that zoneindex has 745 * already adjusted the request size to the appropriate chunk size, which 746 * should optimize our bcopy(). Then copy and return the new pointer. 747 */ 748 if ((nptr = kmalloc(size, type, flags)) == NULL) 749 return(NULL); 750 bcopy(ptr, nptr, min(size, z->z_ChunkSize)); 751 kfree(ptr, type); 752 return(nptr); 753 } 754 755 /* 756 * Return the kmalloc limit for this type, in bytes. 757 */ 758 long 759 kmalloc_limit(struct malloc_type *type) 760 { 761 if (type->ks_limit == 0) { 762 crit_enter(); 763 if (type->ks_limit == 0) 764 malloc_init(type); 765 crit_exit(); 766 } 767 return(type->ks_limit); 768 } 769 770 /* 771 * Allocate a copy of the specified string. 772 * 773 * (MP SAFE) (MAY BLOCK) 774 */ 775 char * 776 kstrdup(const char *str, struct malloc_type *type) 777 { 778 int zlen; /* length inclusive of terminating NUL */ 779 char *nstr; 780 781 if (str == NULL) 782 return(NULL); 783 zlen = strlen(str) + 1; 784 nstr = kmalloc(zlen, type, M_WAITOK); 785 bcopy(str, nstr, zlen); 786 return(nstr); 787 } 788 789 #ifdef SMP 790 /* 791 * free() (SLAB ALLOCATOR) 792 * 793 * Free the specified chunk of memory. 794 */ 795 static 796 void 797 free_remote(void *ptr) 798 { 799 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0); 800 kfree(ptr, *(struct malloc_type **)ptr); 801 } 802 803 #endif 804 805 /* 806 * free (SLAB ALLOCATOR) 807 * 808 * Free a memory block previously allocated by malloc. Note that we do not 809 * attempt to uplodate ks_loosememuse as MP races could prevent us from 810 * checking memory limits in malloc. 811 * 812 * MPSAFE 813 */ 814 void 815 kfree(void *ptr, struct malloc_type *type) 816 { 817 SLZone *z; 818 SLChunk *chunk; 819 SLGlobalData *slgd; 820 struct globaldata *gd; 821 int pgno; 822 823 logmemory_quick(free_beg); 824 gd = mycpu; 825 slgd = &gd->gd_slab; 826 827 if (ptr == NULL) 828 panic("trying to free NULL pointer"); 829 830 /* 831 * Handle special 0-byte allocations 832 */ 833 if (ptr == ZERO_LENGTH_PTR) { 834 logmemory(free_zero, ptr, type, -1, 0); 835 logmemory_quick(free_end); 836 return; 837 } 838 839 /* 840 * Handle oversized allocations. XXX we really should require that a 841 * size be passed to free() instead of this nonsense. 842 * 843 * This code is never called via an ipi. 844 */ 845 { 846 struct kmemusage *kup; 847 unsigned long size; 848 849 kup = btokup(ptr); 850 if (kup->ku_pagecnt) { 851 size = kup->ku_pagecnt << PAGE_SHIFT; 852 kup->ku_pagecnt = 0; 853 #ifdef INVARIANTS 854 KKASSERT(sizeof(weirdary) <= size); 855 bcopy(weirdary, ptr, sizeof(weirdary)); 856 #endif 857 /* 858 * note: we always adjust our cpu's slot, not the originating 859 * cpu (kup->ku_cpuid). The statistics are in aggregate. 860 * 861 * note: XXX we have still inherited the interrupts-can't-block 862 * assumption. An interrupt thread does not bump 863 * gd_intr_nesting_level so check TDF_INTTHREAD. This is 864 * primarily until we can fix softupdate's assumptions about free(). 865 */ 866 crit_enter(); 867 --type->ks_inuse[gd->gd_cpuid]; 868 type->ks_memuse[gd->gd_cpuid] -= size; 869 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) { 870 logmemory(free_ovsz_delayed, ptr, type, size, 0); 871 z = (SLZone *)ptr; 872 z->z_Magic = ZALLOC_OVSZ_MAGIC; 873 z->z_Next = slgd->FreeOvZones; 874 z->z_ChunkSize = size; 875 slgd->FreeOvZones = z; 876 crit_exit(); 877 } else { 878 crit_exit(); 879 logmemory(free_ovsz, ptr, type, size, 0); 880 kmem_slab_free(ptr, size); /* may block */ 881 } 882 logmemory_quick(free_end); 883 return; 884 } 885 } 886 887 /* 888 * Zone case. Figure out the zone based on the fact that it is 889 * ZoneSize aligned. 890 */ 891 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 892 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 893 894 /* 895 * If we do not own the zone then forward the request to the 896 * cpu that does. Since the timing is non-critical, a passive 897 * message is sent. 898 */ 899 if (z->z_CpuGd != gd) { 900 *(struct malloc_type **)ptr = type; 901 #ifdef SMP 902 logmemory(free_request, ptr, type, z->z_ChunkSize, 0); 903 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr); 904 #else 905 panic("Corrupt SLZone"); 906 #endif 907 logmemory_quick(free_end); 908 return; 909 } 910 911 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0); 912 913 if (type->ks_magic != M_MAGIC) 914 panic("free: malloc type lacks magic"); 915 916 crit_enter(); 917 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT; 918 chunk = ptr; 919 920 #ifdef INVARIANTS 921 /* 922 * Attempt to detect a double-free. To reduce overhead we only check 923 * if there appears to be link pointer at the base of the data. 924 */ 925 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) { 926 SLChunk *scan; 927 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) { 928 if (scan == chunk) 929 panic("Double free at %p", chunk); 930 } 931 } 932 chunk_mark_free(z, chunk); 933 #endif 934 935 /* 936 * Put weird data into the memory to detect modifications after freeing, 937 * illegal pointer use after freeing (we should fault on the odd address), 938 * and so forth. XXX needs more work, see the old malloc code. 939 */ 940 #ifdef INVARIANTS 941 if (z->z_ChunkSize < sizeof(weirdary)) 942 bcopy(weirdary, chunk, z->z_ChunkSize); 943 else 944 bcopy(weirdary, chunk, sizeof(weirdary)); 945 #endif 946 947 /* 948 * Add this free non-zero'd chunk to a linked list for reuse, adjust 949 * z_FirstFreePg. 950 */ 951 #ifdef INVARIANTS 952 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd) 953 panic("BADFREE %p", chunk); 954 #endif 955 chunk->c_Next = z->z_PageAry[pgno]; 956 z->z_PageAry[pgno] = chunk; 957 #ifdef INVARIANTS 958 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart) 959 panic("BADFREE2"); 960 #endif 961 if (z->z_FirstFreePg > pgno) 962 z->z_FirstFreePg = pgno; 963 964 /* 965 * Bump the number of free chunks. If it becomes non-zero the zone 966 * must be added back onto the appropriate list. 967 */ 968 if (z->z_NFree++ == 0) { 969 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 970 slgd->ZoneAry[z->z_ZoneIndex] = z; 971 } 972 973 --type->ks_inuse[z->z_Cpu]; 974 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; 975 976 /* 977 * If the zone becomes totally free, and there are other zones we 978 * can allocate from, move this zone to the FreeZones list. Since 979 * this code can be called from an IPI callback, do *NOT* try to mess 980 * with kernel_map here. Hysteresis will be performed at malloc() time. 981 */ 982 if (z->z_NFree == z->z_NMax && 983 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) 984 ) { 985 SLZone **pz; 986 987 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) 988 ; 989 *pz = z->z_Next; 990 z->z_Magic = -1; 991 z->z_Next = slgd->FreeZones; 992 slgd->FreeZones = z; 993 ++slgd->NFreeZones; 994 } 995 logmemory_quick(free_end); 996 crit_exit(); 997 } 998 999 #if defined(INVARIANTS) 1000 /* 1001 * Helper routines for sanity checks 1002 */ 1003 static 1004 void 1005 chunk_mark_allocated(SLZone *z, void *chunk) 1006 { 1007 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 1008 __uint32_t *bitptr; 1009 1010 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex)); 1011 bitptr = &z->z_Bitmap[bitdex >> 5]; 1012 bitdex &= 31; 1013 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk)); 1014 *bitptr |= 1 << bitdex; 1015 } 1016 1017 static 1018 void 1019 chunk_mark_free(SLZone *z, void *chunk) 1020 { 1021 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 1022 __uint32_t *bitptr; 1023 1024 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex)); 1025 bitptr = &z->z_Bitmap[bitdex >> 5]; 1026 bitdex &= 31; 1027 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk)); 1028 *bitptr &= ~(1 << bitdex); 1029 } 1030 1031 #endif 1032 1033 /* 1034 * kmem_slab_alloc() 1035 * 1036 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the 1037 * specified alignment. M_* flags are expected in the flags field. 1038 * 1039 * Alignment must be a multiple of PAGE_SIZE. 1040 * 1041 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), 1042 * but when we move zalloc() over to use this function as its backend 1043 * we will have to switch to kreserve/krelease and call reserve(0) 1044 * after the new space is made available. 1045 * 1046 * Interrupt code which has preempted other code is not allowed to 1047 * use PQ_CACHE pages. However, if an interrupt thread is run 1048 * non-preemptively or blocks and then runs non-preemptively, then 1049 * it is free to use PQ_CACHE pages. 1050 * 1051 * This routine will currently obtain the BGL. 1052 * 1053 * MPALMOSTSAFE - acquires mplock 1054 */ 1055 static void * 1056 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) 1057 { 1058 vm_size_t i; 1059 vm_offset_t addr; 1060 int count, vmflags, base_vmflags; 1061 thread_t td; 1062 1063 size = round_page(size); 1064 addr = vm_map_min(&kernel_map); 1065 1066 /* 1067 * Reserve properly aligned space from kernel_map. RNOWAIT allocations 1068 * cannot block. 1069 */ 1070 if (flags & M_RNOWAIT) { 1071 if (try_mplock() == 0) 1072 return(NULL); 1073 } else { 1074 get_mplock(); 1075 } 1076 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1077 crit_enter(); 1078 vm_map_lock(&kernel_map); 1079 if (vm_map_findspace(&kernel_map, addr, size, align, &addr)) { 1080 vm_map_unlock(&kernel_map); 1081 if ((flags & M_NULLOK) == 0) 1082 panic("kmem_slab_alloc(): kernel_map ran out of space!"); 1083 crit_exit(); 1084 vm_map_entry_release(count); 1085 rel_mplock(); 1086 return(NULL); 1087 } 1088 1089 /* 1090 * kernel_object maps 1:1 to kernel_map. 1091 */ 1092 vm_object_reference(&kernel_object); 1093 vm_map_insert(&kernel_map, &count, 1094 &kernel_object, addr, addr, addr + size, 1095 VM_MAPTYPE_NORMAL, 1096 VM_PROT_ALL, VM_PROT_ALL, 1097 0); 1098 1099 td = curthread; 1100 1101 base_vmflags = 0; 1102 if (flags & M_ZERO) 1103 base_vmflags |= VM_ALLOC_ZERO; 1104 if (flags & M_USE_RESERVE) 1105 base_vmflags |= VM_ALLOC_SYSTEM; 1106 if (flags & M_USE_INTERRUPT_RESERVE) 1107 base_vmflags |= VM_ALLOC_INTERRUPT; 1108 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) 1109 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]); 1110 1111 1112 /* 1113 * Allocate the pages. Do not mess with the PG_ZERO flag yet. 1114 */ 1115 for (i = 0; i < size; i += PAGE_SIZE) { 1116 vm_page_t m; 1117 1118 /* 1119 * VM_ALLOC_NORMAL can only be set if we are not preempting. 1120 * 1121 * VM_ALLOC_SYSTEM is automatically set if we are preempting and 1122 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is 1123 * implied in this case), though I'm not sure if we really need to 1124 * do that. 1125 */ 1126 vmflags = base_vmflags; 1127 if (flags & M_WAITOK) { 1128 if (td->td_preempted) 1129 vmflags |= VM_ALLOC_SYSTEM; 1130 else 1131 vmflags |= VM_ALLOC_NORMAL; 1132 } 1133 1134 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags); 1135 1136 /* 1137 * If the allocation failed we either return NULL or we retry. 1138 * 1139 * If M_WAITOK is specified we wait for more memory and retry. 1140 * If M_WAITOK is specified from a preemption we yield instead of 1141 * wait. Livelock will not occur because the interrupt thread 1142 * will not be preempting anyone the second time around after the 1143 * yield. 1144 */ 1145 if (m == NULL) { 1146 if (flags & M_WAITOK) { 1147 if (td->td_preempted) { 1148 vm_map_unlock(&kernel_map); 1149 lwkt_yield(); 1150 vm_map_lock(&kernel_map); 1151 } else { 1152 vm_map_unlock(&kernel_map); 1153 vm_wait(0); 1154 vm_map_lock(&kernel_map); 1155 } 1156 i -= PAGE_SIZE; /* retry */ 1157 continue; 1158 } 1159 1160 /* 1161 * We were unable to recover, cleanup and return NULL 1162 */ 1163 while (i != 0) { 1164 i -= PAGE_SIZE; 1165 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i)); 1166 /* page should already be busy */ 1167 vm_page_free(m); 1168 } 1169 vm_map_delete(&kernel_map, addr, addr + size, &count); 1170 vm_map_unlock(&kernel_map); 1171 crit_exit(); 1172 vm_map_entry_release(count); 1173 rel_mplock(); 1174 return(NULL); 1175 } 1176 } 1177 1178 /* 1179 * Success! 1180 * 1181 * Mark the map entry as non-pageable using a routine that allows us to 1182 * populate the underlying pages. 1183 * 1184 * The pages were busied by the allocations above. 1185 */ 1186 vm_map_set_wired_quick(&kernel_map, addr, size, &count); 1187 crit_exit(); 1188 1189 /* 1190 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. 1191 */ 1192 for (i = 0; i < size; i += PAGE_SIZE) { 1193 vm_page_t m; 1194 1195 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i)); 1196 m->valid = VM_PAGE_BITS_ALL; 1197 /* page should already be busy */ 1198 vm_page_wire(m); 1199 vm_page_wakeup(m); 1200 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL, 1); 1201 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) 1202 bzero((char *)addr + i, PAGE_SIZE); 1203 vm_page_flag_clear(m, PG_ZERO); 1204 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED)); 1205 vm_page_flag_set(m, PG_REFERENCED); 1206 } 1207 vm_map_unlock(&kernel_map); 1208 vm_map_entry_release(count); 1209 rel_mplock(); 1210 return((void *)addr); 1211 } 1212 1213 /* 1214 * kmem_slab_free() 1215 * 1216 * MPALMOSTSAFE - acquires mplock 1217 */ 1218 static void 1219 kmem_slab_free(void *ptr, vm_size_t size) 1220 { 1221 get_mplock(); 1222 crit_enter(); 1223 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); 1224 crit_exit(); 1225 rel_mplock(); 1226 } 1227 1228